Supercomputing in Plain English Applications and Types of Parallelism PRESENTERNAME PRESENTERTITLE PRESENTERDEPARTMENT PRESENTERINSTITUTION DAY MONTH DATE

SupercomputingSupercomputingin Plain Englishin Plain English

Applications and Applications andTypes of ParallelismTypes of Parallelism

PRESENTERNAMEPRESENTERTITLE

PRESENTERDEPARTMENTPRESENTERINSTITUTION

DAY MONTH DATE YEAR

Your Logo Here

Slides by Henry Neeman, University of Oklahoma

Your Logo Here

Outline Monte Carlo: Client-Server N-Body: Task Parallelism Transport: Data Parallelism

Supercomputing in Plain English: Apps & Par Types INSTITUTION, DAY MONTH DATE YEAR 2

Monte Carlo:Client-Server

[1]

Your Logo Here

Embarrassingly Parallel

An application is known as embarrassingly parallel if its parallel implementation:

1. can straightforwardly be broken up into roughly equal amounts of work per processor, AND

2. has minimal parallel overhead (for example, communication among processors).

We love embarrassingly parallel applications, because they get near-perfect parallel speedup, sometimes with modest programming effort.

Embarrassingly parallel applications are also known as loosely coupled.


Your Logo Here

Monte Carlo Methods

Monte Carlo is a European city where people gamble; that is, they play games of chance, which involve randomness.

Monte Carlo methods are ways of simulating (or otherwise calculating) physical phenomena based on randomness.

Monte Carlo simulations typically are embarrassingly parallel.


Your Logo Here

Monte Carlo Methods: ExampleSuppose you have some physical phenomenon. For example,

consider High Energy Physics, in which we bang tiny particles together at incredibly high speeds.

BANG!We want to know, say, the average properties of this

phenomenon.There are infinitely many ways that two particles can be

banged together.So, we can’t possibly simulate all of them.


Your Logo Here

Monte Carlo Methods: Example

Suppose you have some physical phenomenon. For example, consider High Energy Physics, in which we bang tiny particles together at incredibly high speeds.

BANG!There are infinitely many ways that two particles can be

banged together.So, we can’t possibly simulate all of them.Instead, we can randomly choose a finite subset of these

infinitely many ways and simulate only the subset.


Your Logo Here

Monte Carlo Methods: ExampleSuppose you have some physical phenomenon. For example,

consider High Energy Physics, in which we bang tiny particles together at incredibly high speeds.

BANG!There are infinitely many ways that two particles can be banged

together.We randomly choose a finite subset of these infinitely many

ways and simulate only the subset.The average of this subset will be close to the actual average.


Your Logo Here

Monte Carlo Methods

In a Monte Carlo method, you randomly generate a large number of example cases (realizations) of a phenomenon, and then take the average of the properties of these realizations.

When the average of the realizations converges (that is, doesn’t change substantially if new realizations are generated), then the Monte Carlo simulation stops.


Your Logo Here

MC: Embarrassingly ParallelMonte Carlo simulations are embarrassingly parallel, because

each realization is completely independent of all of the other realizations.

That is, if you’re going to run a million realizations, then:

1. you can straightforwardly break into roughly (Million / Np) chunks of realizations, one chunk for each of the Np processors, AND

2. the only parallel overhead (for example, communication) comes from tracking the average properties, which doesn’t have to happen very often.


Your Logo Here

Serial Monte Carlo (C)

Suppose you have an existing serial Monte Carlo simulation:int main (int argc, char** argv)

{ /* main */ read_input(…); for (realization = 0; realization < number_of_realizations; realization++) { generate_random_realization(…); calculate_properties(…); } /* for realization */ calculate_average(…);} /* main */

How would you parallelize this?


Your Logo Here

Serial Monte Carlo (F90)

Suppose you have an existing serial Monte Carlo simulation:PROGRAM monte_carlo CALL read_input(…) DO realization = 1, number_of_realizations CALL generate_random_realization(…) CALL calculate_properties(…) END DO CALL calculate_average(…)END PROGRAM monte_carlo

How would you parallelize this?


Your Logo Here

Parallel Monte Carlo (C)int main (int argc, char** argv){ /* main */ [MPI startup] if (my_rank == server_rank) { read_input(…); } mpi_error_code = MPI_Bcast(…); for (realization = 0; realization < number_of_realizations; realization++) { generate_random_realization(…); calculate_realization_properties(…); calculate_local_running_average(...); } /* for realization */ if (my_rank == server_rank) { [collect properties] } else { [send properties] } calculate_global_average_from_local_averages(…) output_overall_average(...) [MPI shutdown]} /* main */


Your Logo Here

Parallel Monte Carlo (F90)PROGRAM monte_carlo [MPI startup] IF (my_rank == server_rank) THEN CALL read_input(…) END IF CALL MPI_Bcast(…) DO realization = 1, number_of_realizations CALL generate_random_realization(…) CALL calculate_realization_properties(…) CALL calculate_local_running_average(...) END DO IF (my_rank == server_rank) THEN [collect properties] ELSE [send properties] END IF CALL calculate_global_average_from_local_averages(…) CALL output_overall_average(...) [MPI shutdown]END PROGRAM monte_carlo


N-Body:Task Parallelism and

Collective Communication

[2]

Your Logo Here

N Bodies


Your Logo Here

N-Body ProblemsAn N-body problem is a problem involving N “bodies” –

that is, particles (for example, stars, atoms) – each of which applies a force to all of the others.

For example, if you have N stars, then each of the N stars exerts a force (gravity) on all of the other N–1 stars.

Likewise, if you have N atoms, then every atom exerts a force (nuclear) on all of the other N–1 atoms.


Your Logo Here

1-Body Problem

When N is 1, you have a simple 1-Body Problem: a single particle, with no forces acting on it.

Given the particle’s position P and velocity V at some time t0, you can trivially calculate the particle’s position at time t0+Δt:

P(t0+Δt) = P(t0) + VΔt

V(t0+Δt) = V(t0)


Your Logo Here

2-Body ProblemWhen N is 2, you have – surprise! – a 2-Body Problem:

exactly 2 particles, each exerting a force that acts on the other.

The relationship between the 2 particles can be expressed as a differential equation that can be solved analytically, producing a closed-form solution.

So, given the particles’ initial positions and velocities, you can trivially calculate their positions and velocities at any later time.


Your Logo Here

3-Body ProblemWhen N is 3, you have – surprise! – a 3-Body Problem:

exactly 3 particles, each exerting a force that acts on the other 2.

The relationship between the 3 particles can be expressed as a differential equation that can be solved using an infinite series, producing a closed-form solution, due to Karl Fritiof Sundman in 1912.

However, in practice, the number of terms of the infinite series that you need to calculate to get a reasonable solution is so large that the infinite series is impractical, so you’re stuck with the generalized formulation.


Your Logo Here

N-Body Problems (N > 3)When N > 3, you have a general N-Body Problem: N particles,

each exerting a force that acts on the other N-1 particles.The relationship between the N particles can be expressed as a

differential equation that can be solved using an infinite series, producing a closed-form solution, due to Qiudong Wang in 1991.

However, in practice, the number of terms of the infinite series that you need to calculate to get a reasonable solution is so large that the infinite series is impractical, so you’re stuck with the generalized formulation.


Your Logo Here

N-Body Problems (N > 3)

For N greater than 3, the relationship between the N particles can be expressed as a differential equation that can be solved using an infinite series, producing a closed-form solution, but convergence takes so long that this approach is impractical.

So, numerical simulation is pretty much the only way to study groups of 3 or more bodies.

Popular applications of N-body codes include: astronomy (that is, galaxy formation, cosmology); chemistry (that is, protein folding, molecular dynamics).

Note that, for N bodies, there are on the order of N2 forces, denoted O(N2).


Your Logo Here

N Bodies


Your Logo Here

Force #1


A

Your Logo Here

Force #2


A

Your Logo Here

Force #3


A

Your Logo Here

Force #4


A

Your Logo Here

Force #5


A

Your Logo Here

Force #6


A

Your Logo Here

Force #N-1


A

Your Logo Here

N-Body Problems

Given N bodies, each body exerts a force on all of the other N – 1 bodies.

Therefore, there are N • (N – 1) forces in total.

You can also think of this as (N • (N – 1)) / 2 forces, in the sense that the force from particle A to particle B is the same (except in the opposite direction) as the force from particle B to particle A.


Your Logo Here

Aside: Big-O Notation

Let’s say that you have some task to perform on a certain number of things, and that the task takes a certain amount of time to complete.

Let’s say that the amount of time can be expressed as a polynomial on the number of things to perform the task on.

For example, the amount of time it takes to read a book might be proportional to the number of words, plus the amount of time it takes to settle into your favorite easy chair.

C1 . N + C2


Your Logo Here

Big-O: Dropping the Low TermC1 . N + C2

When N is very large, the time spent settling into your easy chair becomes such a small proportion of the total time that it’s virtually zero.

So from a practical perspective, for large N, the polynomial reduces to:

C1 . NIn fact, for any polynomial, if N is large, then all of the terms

except the highest-order term are irrelevant.


Your Logo Here

Big-O: Dropping the ConstantC1 . N

Computers get faster and faster all the time. And there are many different flavors of computers, having many different speeds.

So, computer scientists don’t care about the constant, only about the order of the highest-order term of the polynomial.

They indicate this with Big-O notation:O(N)

This is often said as: “of order N.”


Your Logo Here

N-Body Problems

Given N bodies, each body exerts a force on all of the other N – 1 bodies.

Therefore, there are N • (N – 1) forces total.In Big-O notation, that’s O(N2) forces.

So, calculating the forces takes O(N2) time to execute.

But, there are only N particles, each taking up the same amount of memory, so we say that N-body codes are of:

O(N) spatial complexity (memory) O(N2) time complexity


Your Logo Here

O(N2) Forces


Note that this picture shows only the forces between A and everyone else.

A

Your Logo Here

How to Calculate?

Whatever your physics is, you have some function, F(Bi,Bj), that expresses the force between two bodies Bi and Bj.

For example, for stars and galaxies,

F(A,B) = G · mBi · mBj

/ dist(Bi, Bj)2

where G is the gravitational constant and m is the mass of the body in question.

If you have all of the forces for every pair of particles, then you can calculate their sum, obtaining the force on every particle.

From that, you can calculate every particle’s new position and velocity.


Your Logo Here

How to Parallelize?

Okay, so let’s say you have a nice serial (single-CPU) code that does an N-body calculation.

How are you going to parallelize it?You could: have a server feed particles to processes; have a server feed interactions to processes; have each process decide on its own subset of the particles,

and then share around the forces; have each process decide its own subset of the interactions,

and then share around the forces.


Your Logo Here

Do You Need a Server?

Let’s say that you have N bodies, and therefore you have ½ N (N - 1) interactions (every particle interacts with all of the others, but you don’t need to calculate both Bi Bj and Bj Bi).

Do you need a server?

Well, can each processor determine, on its own, either (a) which of the bodies to process, or (b) which of the interactions to process?

If the answer is yes, then you don’t need a server.


Your Logo Here

Parallelize How?

Suppose you have Np processors.

Should you parallelize: by assigning a subset of N / Np of the bodies to each

processor, OR by assigning a subset of ½ N (N - 1) / Np of the interactions

to each processor?


Your Logo Here

Data vs. Task Parallelism Data Parallelism means parallelizing by giving a subset of

the data to each process, and then each process performs the same tasks on the different subsets of data.

Task Parallelism means parallelizing by giving a subset of the tasks to each process, and then each process performs a different subset of tasks on the same data.


Your Logo Here

Data Parallelism for N-Body?If you parallelize an N-body code by data, then each processor

gets N / Np pieces of data.For example, if you have 8 bodies and 2 processors, then: Processor P0 gets the first 4 bodies; Processor P1 gets the second 4 bodies.But, every piece of data (that is, every body) has to interact

with every other piece of data, to calculate the forces.So, every processor will have to send all of its data to all of the

other processors, for every single interaction that it calculates.

That’s a lot of communication!


Your Logo Here

Task Parallelism for N-body?If you parallelize an N-body code by task, then each processor

gets all of the pieces of data that describe the particles (for example, positions, velocities, masses).

Then, each processor can calculate its subset of the interaction forces on its own, without talking to any of the other processors.

But, at the end of the force calculations, everyone has to share all of the forces that have been calculated, so that each particle ends up with the total force that acts on it (global reduction).


Your Logo Here

MPI_Reduce (C)Here’s the C syntax for MPI_Reduce: mpi_error_code = MPI_Reduce(sendbuffer, recvbuffer, count, datatype, operation, root, communicator, mpi_error_code);For example, to do a sum over all of the particle forces: mpi_error_code = MPI_Reduce( local_particle_force_sum, global_particle_force_sum, number_of_particles, MPI_DOUBLE, MPI_SUM, server_process, MPI_COMM_WORLD);


Your Logo Here

MPI_Reduce (F90)Here’s the Fortran 90 syntax for MPI_Reduce: CALL MPI_Reduce(sendbuffer, recvbuffer, & & count, datatype, operation, & & root, communicator, mpi_error_code)For example, to do a sum over all of the particle forces: CALL MPI_Reduce( & & local_particle_force_sum, & & global_particle_force_sum, & & number_of_particles, & & MPI_DOUBLE_PRECISION, MPI_SUM, & & server_process, MPI_COMM_WORLD, & & mpi_error_code)


Your Logo Here

Sharing the ResultIn the N-body case, we don’t want just one processor to know

the result of the sum, we want every processor to know.So, we could do a reduce followed immediately by a

broadcast.

But, MPI gives us a routine that packages all of that for us: MPI_Allreduce.

MPI_Allreduce is just like MPI_Reduce except that every process gets the result (so we drop the server_process argument).


Your Logo Here

MPI_Allreduce (C)Here’s the C syntax for MPI_Allreduce: mpi_error_code = MPI_Allreduce( sendbuffer, recvbuffer, count, datatype, operation, communicator);For example, to do a sum over all of the particle forces: mpi_error_code = MPI_Allreduce( local_particle_force_sum, global_particle_force_sum, number_of_particles, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);


Your Logo Here

MPI_Allreduce (F90)Here’s the Fortran 90 syntax for MPI_Allreduce: CALL MPI_Allreduce( & & sendbuffer, recvbuffer, count, & & datatype, operation, & & communicator, mpi_error_code)For example, to do a sum over all of the particle forces: CALL MPI_Allreduce( & & local_particle_force_sum, & & global_particle_force_sum, & & number_of_particles, & & MPI_DOUBLE_PRECISION, MPI_SUM, &

& MPI_COMM_WORLD, mpi_error_code)


Your Logo Here

Collective Communications

A collective communication is a communication that is shared among many processes, not just a sender and a receiver.

MPI_Reduce and MPI_Allreduce are collective communications.

Others include: broadcast, gather/scatter, all-to-all.


Your Logo Here

Collectives Are Expensive

Collective communications are very expensive relative to point-to-point communications, because so much more communication has to happen.

But, they can be much cheaper than doing zillions of point-to-point communications, if that’s the alternative.


Transport:Data Parallelism

[2]

Your Logo Here

What is a Simulation?

All physical science ultimately is expressed as calculus (for example, differential equations).

Except in the simplest (uninteresting) cases, equations based on calculus can’t be directly solved on a computer.

Therefore, all physical science on computers has to be approximated.


Your Logo Here

I Want the Area Under This Curve!


How can I get the area under this curve?

Your Logo Here

A Riemann Sum


Δx

{

yi

Area under the curve ≈ ∑=

Δn

ii xy

1

Ceci n’est pas un area under the curve: it’s approximate!

[3]

Your Logo Here

A Riemann Sum


Δx

{

yi


Δn

ii xy

1

Ceci n’est pas un area under the curve: it’s approximate!

Your Logo Here

A Better Riemann Sum


Δx

{

yi


Δn

ii xy

1

More, smaller rectangles produce a better approximation.

Your Logo Here

The Best Riemann Sum


Area under the curve = ∑ ∫∞

=

≡1i

i ydxdxy

In the limit, infinitely many infinitesimally small rectangles produce the exact area.

Your Logo Here

The Best Riemann Sum


Area under the curve = ∑ ∫∞

=

≡1i

i ydxdxy

In the limit, infinitely many infinitesimally small rectangles produce the exact area.

Your Logo Here

Differential Equations

A differential equation is an equation in which differentials (for example, dx) appear as variables.

Most physics is best expressed as differential equations.Very simple differential equations can be solved in “closed

form,” meaning that a bit of algebraic manipulation gets the exact answer.

Interesting differential equations, like the ones governing interesting physics, can’t be solved in close form.

Solution: approximate!


Your Logo Here

A Discrete Mesh of Data


Data live

here!

Your Logo Here

A Discrete Mesh of Data


Data live

here!

Your Logo Here

Finite Difference

A typical (though not the only) way of approximating the solution of a differential equation is through finite differencing: convert each dx (infinitely thin) into a Δx (has finite nonzero width).


Your Logo Here

Navier-Stokes Equation


⎥⎥⎦

⎤

⎢⎢⎣

⎡⋅∇+⎟

⎟⎠

⎞⎜⎜⎝

⎛

∂∂

+∂∂

∂∂

= uiji

j

j

i

j

i

x

u

x

u

xV

Fλδη

⎥⎥⎦

⎤

⎢⎢⎣

⎡⋅∇+⎟

⎟⎠

⎞⎜⎜⎝

⎛

ΔΔ

+ΔΔ

ΔΔ

= uiji

j

j

i

j

i

x

u

x

u

xV

Fλδη

Differential Equation

Finite Difference Equation

The Navier-Stokes equations governs the movement of fluids (water, air, etc).

Your Logo Here

Cartesian Coordinates


x

y

Your Logo Here

Structured Mesh

A structured mesh is like the mesh on the previous slide. It’s nice and regular and rectangular, and can be stored in a standard Fortran or C or C++ array of the appropriate dimension and shape.

REAL,DIMENSION(nx,ny,nz) :: u

float u[nx][ny][nz];


Your Logo Here

Flow in Structured MeshesWhen calculating flow in a structured mesh, you typically use

a finite difference equation, like so: unewi,j = F(Δt, uoldi,j, uoldi-1,j, uoldi+1,j, uoldi,j-1, uoldi,j+1)

for some function F, where uoldi,j is at time t and unewi,j is at time t + Δt.

In other words, you calculate the new value of ui,j, based on its old value as well as the old values of its immediate neighbors.

Actually, it may use neighbors a few farther away.


Your Logo Here

Ghost Boundary Zones


Your Logo Here

Ghost Boundary Zones

We want to calculate values in the part of the mesh that we care about, but to do that, we need values on the boundaries.

For example, to calculate unew1,1, you need uold0,1 and uold1,0.

Ghost boundary zones are mesh zones that aren’t really part of the problem domain that we care about, but that hold boundary data for calculating the parts that we do care about.


Your Logo Here

Using Ghost Boundary Zones (C)

A good basic algorithm for flow that uses ghost boundary zones is:

for (timestep = 0; timestep < number_of_timesteps; timestep++) { fill_ghost_boundary(…); advance_to_new_from_old(…);}This approach generally works great on a serial code.


Your Logo Here

Using Ghost Boundary Zones (F90)

A good basic algorithm for flow that uses ghost boundary zones is:

DO timestep = 1, number_of_timesteps CALL fill_ghost_boundary(…) CALL advance_to_new_from_old(…)END DOThis approach generally works great on a serial code.


Your Logo Here

Ghost Boundary Zones in MPI

What if you want to parallelize a Cartesian flow code in MPI?

You’ll need to: decompose the mesh into submeshes; figure out how each submesh talks to its neighbors.


Your Logo Here

Data Decomposition


Your Logo Here

Data Decomposition

We want to split the data into chunks of equal size, and give each chunk to a processor to work on.

Then, each processor can work independently of all of the others, except when it’s exchanging boundary data with its neighbors.


Your Logo Here

MPI_Cart_*

MPI supports exactly this kind of calculation, with a set of functions MPI_Cart_*:

MPI_Cart_create MPI_Cart_coords MPI_Cart_shift

These routines create and describe a new communicator, one that replaces MPI_COMM_WORLD in your code.


Your Logo Here

MPI_SendrecvMPI_Sendrecv is just like an MPI_Send followed by an MPI_Recv, except that it’s much better than that.

With MPI_Send and MPI_Recv, these are your choices: Everyone calls MPI_Recv, and then everyone calls MPI_Send.

Everyone calls MPI_Send, and then everyone calls MPI_Recv.

Some call MPI_Send while others call MPI_Recv, and then they swap roles.


Your Logo Here

Why not Recv then Send?

Suppose that everyone calls MPI_Recv, and then everyone calls MPI_Send.

MPI_Recv(incoming_data, ...); MPI_Send(outgoing_data, ...);

Well, these routines are blocking, meaning that the communication has to complete before the process can continue on farther into the program.

That means that, when everyone calls MPI_Recv, they’re waiting for someone else to call MPI_Send.

We call this deadlock.

Officially, the MPI standard guarantees that THIS APPROACH WILL ALWAYS FAIL.


Your Logo Here

Why not Send then Recv?

Suppose that everyone calls MPI_Send, and then everyone calls MPI_Recv:

MPI_Send(outgoing_data, ...); MPI_Recv(incoming_data, ...);

Well, this will only work if there’s enough buffer space available to hold everyone’s messages until after everyone is done sending.

Sometimes, there isn’t enough buffer space.

Officially, the MPI standard allows MPI implementers to support this, but it isn’t part of the official MPI standard; that is, a particular MPI implementation doesn’t have to allow it, so THIS WILL SOMETIMES FAIL.


Your Logo Here

Alternate Send and Recv?Suppose that some processors call MPI_Send while others

call MPI_Recv, and then they swap roles:

if ((my_rank % 2) == 0) { MPI_Send(outgoing_data, ...); MPI_Recv(incoming_data, ...); } else { MPI_Recv(incoming_data, ...); MPI_Send(outgoing_data, ...); }This will work, and is sometimes used, but it can be painful to

manage – especially if you have an odd number of processors.


Your Logo Here

MPI_Sendrecv

MPI_Sendrecv allows each processor to simultaneously send to one processor and receive from another.

For example, P1 could send to P0 while simultaneously receiving from P2 .

(Note that the send and receive don’t have to literally be simultaneous, but we can treat them as so in writing the code.)

This is exactly what we need in Cartesian flow: we want the boundary data to come in from the east while we send boundary data out to the west, and then vice versa.

These are called shifts.


Your Logo Here

MPI_Sendrecv mpi_error_code =

MPI_Sendrecv( westward_send_buffer, westward_send_size, MPI_REAL, west_neighbor_process, westward_tag, westward_recv_buffer, westward_recv_size, MPI_REAL, east_neighbor_process, westward_tag, cartesian_communicator, mpi_status);

This call sends to west_neighbor_process the data in westward_send_buffer, and at the same time receives from east_neighbor_process a bunch of data that end up in westward_recv_buffer.


Your Logo Here

Why MPI_Sendrecv?

The advantage of MPI_Sendrecv is that it allows us the luxury of no longer having to worry about who should send when and who should receive when.

This is exactly what we need in Cartesian flow: we want the boundary information to come in from the east while we send boundary information out to the west – without us having to worry about deciding who should do what to who when.


Your Logo Here

MPI_Sendrecv


Conceptin Principle

Conceptin practice

Your Logo Here

MPI_Sendrecv


Conceptin practice

westward_send_buffer westward_recv_buffer

ActualImplementation

Your Logo Here

What About Edges and Corners?

If your numerical method involves faces, edges and/or corners, don’t despair.

It turns out that, if you do the following, you’ll handle those correctly:

When you send, send the entire ghost boundary’s worth, including the ghost boundary of the part you’re sending.

Do in this order: all east-west; all north-south; all up-down.

At the end, everything will be in the correct place.


Thanks for your attention!

Questions?

Your Logo Here

References


[1] http://en.wikipedia.org/wiki/Monte_carlo_simulation[2] http://en.wikipedia.org/wiki/N-body_problem[3] http://lostbiro.com/blog/wp-content/uploads/2007/10/Magritte-Pipe.jpg

Documents

Supercomputing in Plain English Applications and Types of Parallelism PRESENTERNAME PRESENTERTITLE PRESENTERDEPARTMENT PRESENTERINSTITUTION DAY MONTH DATE